Distributed design optimization of multi-component systems using meta models and topology optimization

• Published in: Structural and multidisciplinary optimization Vol. 67; no. 9; p. 160
• Main Authors: Krischer, Lukas, Endress, Felix, Wanninger, Tobias, Zimmermann, Markus
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2024; Springer Nature B.V
• Subjects: Computational Mathematics and Numerical Analysis; Coordination; Decomposition; Design optimization; Engineering; Engineering Design; Parallel degrees of freedom; Research Paper; Stiffness matrix; Theoretical and Applied Mechanics; Topology optimization; Distributed design optimization; Topology optimization; Machine learning
• ISSN: 1615-147X; 1615-1488; 1615-1488
• DOI: 10.1007/s00158-024-03836-5

Distributed optimization architectures decompose large monolithic optimization problems into sets of smaller and more manageable optimization subproblems. To ensure consistency and convergence towards a global optimum, however, cumbersome coordination is necessary and often not sufficient. A distributed optimization architecture was previously proposed that does not require coordination. This so-called Informed Decomposition is based on two types of optimization problems: (1) one for system optimization to produce stiffness requirements on components using pre-trained meta models and (2) one for the optimization of components with two interfaces to produce detailed geometries that satisfy the stiffness requirements. Each component optimization problem can be carried out independently and in parallel. This paper extends the approach to three-dimensional structures consisting of components with six degrees of freedom per interface, thus significantly increasing the applicability to practical problems. For this, an 8-dimensional representation of the general 12 x 12 interface stiffness matrix for components is derived. Meta models for mass estimation and physical feasibility of stiffness targets are trained using an active-learning strategy. A simple two-component structure and a large robot structure consisting of four components subject to constraints for 100 different loading scenarios are optimized. The example results are at most 12.9% heavier than those of a monolithic optimization.